BTGAN: Training GAN with Balanced Triplet Loss and Two-Branch Architecture

Abstract

TripletGAN is a variant of Generative Adversarial Network (GAN) by replacing the classification loss of discriminator with a triplet loss. Although TripletGAN delivers better mode coverage than vanilla GAN thanks to the characteristics of adversarial triplet loss that maximizes the embedding distance between generated samples, its adversarial training method suffers from the drawback that some generated images tend to deviate from the real sample distribution and noisy images are produced as we increase the number of iterations of training. In this paper, we propose an adversarially balanced triplet loss with four dynamic coefficients to achieve a trade-off between the quality and the diversity of generated samples. We also design a novel network architecture to provide GANs with an auto-encoding ability. Extensive experiments demonstrate the effectiveness of our proposed methods in terms of alleviating the problem in TripletGAN and the superiority in terms of reconstruction over some methods that directly train generator and encoder such as O-GAN.